Fresh details from Chinese memory maker ChangXin Memory Technologies, or CXMT, around a new DRAM architecture have stirred fresh worries: more supply could show up and hit pricing. But Peter Lee (이세철), a global markets semiconductor analyst at Citigroup, said in an interview that this should not be taken as proof that a wave of cheap products is about to pour into the market. There is still a real gap, he said, between a technology announcement and actual mass production. And in his view, the bigger variable for chip investment right now is simpler: can AI companies keep raising the money needed for data centers and compute purchases?
That lines up pretty closely with remarks made recently by Phison CEO K.S. Pua. ChainNews had earlier reported that Pua said CXMT still needs time to get from validation to yield improvement and then into formal mass production, while Phison’s more immediate problem is material shortages that could affect deliveries. He summed up next year’s memory market like this: "Everything is in short supply except orders."
Citi’s read on CXMT’s 4F²: smaller cells do not settle the cost question
The 4F² idea at the center of this debate is about DRAM cell area and layout. Against the more widely used 6F² structure, squeezing more cells into the same space could lift density and cut bit cost.
Lee said 4F² is not something CXMT alone has pursued. Samsung, SK hynix and Micron have all built related research over time. Whether it gets adopted, he said, comes down to two things: how far existing process technology can still scale, and how this newer architecture stacks up on cost and manufacturing difficulty.
He also said CXMT’s method combines multiple wafers. That means more wafer usage, more processing time, and higher cost. So any gain from shrinking cell area has to be weighed against those added expenses before anyone can judge whether it is truly competitive. For that reason, he does not buy the argument that this technology update should automatically be read as the opening shot of a cheap-memory flood. At the same time, limits on access to advanced equipment are pushing Chinese manufacturers to search harder for alternate routes. That, in turn, keeps pressure on incumbent suppliers to keep spending on research and development.
Weak consumer devices and AI shortages can exist at the same time
Lee said memory demand no longer looks like it used to. Smartphones were once the main growth driver. Now the new demand is far more concentrated in AI data centers, including HBM, server DRAM, SoCAMM2 memory modules and enterprise SSDs. Those products also come with stricter demands on performance, quality and reliability. His point: that still gives leading suppliers room to hold their ground.
Consumer electronics face a different squeeze. When memory prices rise, phones, PCs and game consoles get more expensive to build, and that may push brands to change specifications, pricing or purchase volumes. Weak consumer sales and difficulty getting enough memory for data centers are not opposites, he said. Both can happen. At the same time.
Lee also pointed out that China is still building data centers at home. As local suppliers raise output, local demand could rise too. He said the effect of Chinese capacity expansion should be judged by product categories, yields and customer adoption, not just raw output numbers.
Order quality depends on whether long-term deals carry money behind them
Another worry in the market is whether customers are placing big orders with several memory suppliers at once just to lock in supply, which could make real demand look bigger than it is. Lee said the first thing to check is whether long-term agreements are backed by actual financial commitments.
In the interview, he pointed to arrangements such as prepayments, deposits and escrow accounts. If canceling a contract means taking a financial loss, then double-ordering gets a lot more expensive.
He used restaurant reservations as the example. If booking costs nothing, people may hold several tables at once. But if every restaurant wants payment upfront, choices get tighter and more disciplined. Long-term agreements do not erase the business cycle. Still, a purchase commitment tied to payment constraints is not the same as a demand estimate that can be pulled back at any moment. In his view, the cancellation terms and funding setup in each company’s contracts still have to be judged through formal disclosures.
For AI capex, the question is whether companies can actually raise the money
Lee’s main short-term concern is financing. Straight up. If model companies cannot raise funds on time, or if borrowing costs climb for data center operators, then planned compute expansion could be pushed back. And that would ripple through the supply chain, affecting purchases of GPUs, HBM, server DRAM and storage equipment.
He said the market is really asking whether AI companies can keep securing enough capital. Even if application demand keeps growing, investment plans can still be revised if the money does not show up. That is why he treats real fundraising results and the spending capacity of major technology companies as key indicators for memory demand. Announced investment plans do not turn into supplier revenue by themselves. Financing, construction and equipment delivery still need to happen.
Fab completion is not the same as immediate supply
On the risk that expanding memory capacity could tip the market into oversupply, Lee said the completion of a new fab is only one preparatory step. Tool installation, test runs and yield ramp all take time.
He estimated that some new facilities may need about a year and a half from opening before effective capacity is released in a meaningful way, so a plant’s completion date should not be treated as the point when large product volumes hit the market.
He also said that lower memory configurations in some AI systems should be read as a substitution arrangement caused by limited supply. Customers may accept lower specifications to get systems deployed sooner, and that does not necessarily mean demand for higher-capacity configurations has vanished.
Continual learning could add to HBM and SSD demand
Looking out to 2027, Lee flagged continual learning as a possible new driver for memory demand. In his framework, AI first boosts HBM demand through model training. Then wider inference lifts demand for server memory and storage. If models start updating their capabilities using data generated during use, then data traffic between training and inference would become more frequent.
HBM provides the high bandwidth needed for compute, while SSDs handle data storage and some KV Cache data linked to inference workloads. Lee said HDDs and SSDs overlap in some use cases, but they are different in speed and deployment scenarios, so an increase in HDD production should not be read as proof that SSD demand will be widely replaced.
In the interview, Lee estimated that HBM demand could rise about 60% in 2027 and get close to 70% in 2028. He said those numbers are his demand estimates, not supplier revenue guidance. Over a longer stretch, he added, cooperation among AI agents could extend computing demand into round-the-clock task execution.